The Ripple Network is a multi-detached project consisting of two investigations. The first, El Xarxaire Pabellón, is a form-found waffle shell pavilion for Plaza del Mar, Barceloneta an exposed heat-island space with a critical lack of shade. Wallacei multi-objective optimisation evolves the canopy form across three objectives: solar radiation, direct sun hours and open views. A Decision Tree Model extracts feature importance scores as data-driven TOPSIS weights, ranking 400 configurations across 20 generations. Cluster analysis selects the optimal result. The second investigation shifts from performance to production: evaluating the most buildable 3D-printed tile layout by analysing cost, number of printed pieces, module sizes and fabrication time bridging the gap between the best-performing design and what can realistically be made.
For this assignment we started the optimization by building on the pavilion we developed in the Computational Design II Seminar — El Xarxaire, a form-found waffle shell on Plaza del Mar in Barceloneta, generated through Kangaroo physics simulation in Grasshopper. The project offers enough architectural complexity to meaningfully apply this seminar’s methods: a doubly-curved, structurally form-found mesh sitting on a large public plaza, which gives us a geometry to optimize and evaluate.
We connect this to our Introduction to Robotic Fabrication seminar tiles B. the Wave we have 3d printed. Instead of producing the final prototype with OpenNest as a flat-packed waffle structure, we replace that fabrication strategy with the 3D-printed tile system developed in the robotics seminar, an angle-graded module catalogue applied across the pavilion skin. This lets us treat the pavilion surface as a shading system and apply computational optimization to the tile layout.
System Deconstruction
The geometry generation process involves converting flat 2D shapes into a workable mesh, guided by MD Sliders that determine UV coordinates on a base canvas for the pavilion’s organic outer perimeter and interior column placement. Inner support dimensions are adjusted with number sliders for circular column radii (1.0 and 0.708) and rectangular column bounds (3.2 x 1.45). A QuadRemesh component then creates the topological grid with a Target Count of 1500, essential for maintaining mesh density, ensuring sufficient vertices for smooth physics simulation while avoiding computational overload.
Once the QuadRemesh establishes the foundational topology, the static geometry must be translated into a dynamic spring network for simulation. The unified mesh is deconstructed to extract its constituent edges, which act as the structural springs. Simultaneously, the mesh’s naked edges are isolated to define the pavilion’s exterior boundary curves and the interior column voids. The sorting and isolation process ensures that when the structural tension and gravitational forces are applied, the base perimeter and column supports remain physically grounded while the interior mesh faces are free to deform.
The script’s mathematical engine revolves around the form-finding process involving Kangaroo physics and a Graph Mapper to shape the pavilion in 3D. The pavilion’s outer edge rhythm is influenced by Wave Frequency, controlled by adjusting a Sine wave in the Graph Mapper, affecting the number of arches or entrances. A Remap component translates the wave into height, with its Target slider at 6.95 representing the Wave Amplitude for maximum Z-axis elevation. The physics solver manages geometry through Membrane Tension (with an EdgeLengths factor of 0.42 for fabric tightness) and Gravity (using a Load factor of -0.081 to create structural valleys).
Workflow Diagram (Pseudocode)
The project applies Multi-Objective Optimisation with Wallacei, used to evolve the pavilion form for environmental and spatial performance running three competing fitness objectives: minimising incident solar radiation on the surface, minimising direct sun hours.
The Pareto front generated by Wallacei is exported and interpreted using a supervised regression Decision Tree Model developed with the support of AI tools, which identifies which design variables most strongly drive each performance objective and extracts feature importance scores from the data. These importance scores are then used as the basis for the weights in TOPSIS ensuring that the ranking of configurations reflects the actual behaviour of the design system rather than subjective assumptions. AI-generated diagrams are produced to compare the results visually, making the trade-offs between configurations ideal for both technical and non-technical audiences.
Tile application with cluster analysis completes the fabrication phase. The optimised grid mesh is populated with the 3D-printed tile system developed in the Robotic Fabrication seminar, applying the tile texture directly to the surface properties of the pavilion skin. Cluster analysis of the Wallacei population is then used to group solutions by behavioural similarity, allowing the best-performing tile configuration to be selected not as an isolated data point but as a representative of a family of high-performing outcomes — providing a more robust and designerly basis for the final fabrication decision.
Computational Method
Method Objective
Reaching this point required resolving a fundamental technical challenge. The Kangaroo physics simulation used to generate the form-found mesh was initially running on the Bouncy Solver, which executes iteratively in real time and resets unpredictably when Wallacei modifies the genome between individuals. This made multi-objective optimisation impossible the solver could not deliver a stable, converged mesh consistently for every evaluated solution. Switching to the Zombie Solver resolved this, as it runs a fixed number of iterations on demand and outputs a fully converged mesh deterministically, making it compatible with Wallacei’s generational evaluation cycle. This change was essential to producing valid and comparable fitness values across the entire optimised population.
The decision tree then addresses a follow-on interpretive question: across all evaluated solutions, which design variables most strongly determine shading and view performance? The feature-importance output turns a cloud of Pareto solutions into readable rules, explaining the logic behind the optimiser’s choices.
WSM and TOPSIS address the fabrication question: given the optimised form, which tile configuration defined by module size, piece count, cost, and print time represents the best buildable outcome? By ranking alternatives against weighted criteria and cross-checking the two methods against each other, the most feasible solution is identified with evidence rather than assumption.
Fabrication Parameters
Optimal Solution Selection Catalogue
Decision Outcome Criteria
Analysis and Results
Design Outcome and Visualization
